Sales partition for forecasting into textile distribution network

This paper presents a new method for the management of buffer-stocks for textile items by a sales partition approach. The proposed algorithm enables to partition the set of items into different optimal classes, so that items belonging to each class follow the same forecasting sales graph. This method uses a sequence of merging of classes. In this sequence, an internal index measuring the compactness inside each class and the separability between different classes enables to determine the best partition. Then, for each class created, the compactness is compared to the initial compacteness criterion given by the industrial manager and, the class which doesn't respect the classification accuracy is rejected. A faster searching of forecasted sales graph for an item is thus achieved by the determination of a mean graph corresponding to its family sale.